The Objection an AI Draft Leaves Out: Why Judges Must Check Rulings Against the Original Record Commentary
Instant Images
The Objection an AI Draft Leaves Out: Why Judges Must Check Rulings Against the Original Record

A judgment can contain authentic citations, accurate statements of law and fluent reasoning while failing to answer the objection that could defeat its conclusion. That is a distinctive risk when a court uses artificial intelligence to summarise a file or help prepare text. Checking whether cited authorities exist is necessary; it does not establish that the court has answered the right question. A useful measure of judicial oversight is whether the final reasons can be traced back to the parties’ potentially decisive objections, including those the automated summary left out.

On 9 October 2026, the judiciary of England and Wales published a position statement on the judicial use of AI. It insists that judicial decisions remain with judges, that AI cannot displace evaluation of evidence or judicial reasoning, and that every AI output used in judicial work must be reviewed, verified and approved by the responsible judicial office holder. The statement describes cautious testing of lower-risk applications, including summarisation, chronologies and document analysis. It does not give blanket approval for AI to draft judgments. Its commitments nevertheless pose a concrete design question: what should a judge verify when the danger is not a false quotation, but a missing argument?

The danger is an omission, not merely an invention

Consider a hypothetical civil claim in a jurisdiction where service of a prescribed notice is a prerequisite to possession. The claimant establishes title. The defendant puts service in issue and identifies evidence suggesting the notice never arrived. A draft correctly sets out the law of ownership, cites genuine decisions and grants possession. Yet, if service is a necessary condition of relief and remains disputed, the reasoning has bypassed a question on which the outcome may turn.

Nothing in this example shows that an AI system committed an actual error. The problem is structural. A case summary may omit the objection; a drafting instruction may frame the case exclusively as a title dispute; a reviewer may then validate the authorities mentioned in the resulting text. Each stage can be internally consistent with the previous one, while the original omission survives. Comparing a draft only with the summary from which it was generated therefore offers an inadequate safeguard.

The failure is not cured by a formula such as “all submissions have been considered”, nor by adding a longer discussion of general doctrine. What is needed is a finding about service, an explanation of its legal irrelevance, or another reasoned answer capable of supporting the order actually made. If the point is decisive, polished language is no substitute for adjudication.

The legal standard is selective, but demanding

The European Court of Human Rights has long drawn a distinction between an obligation to give reasons and a duty to answer every argument separately. In Ruiz Torija v Spain, decided on 9 December 1994, a clearly raised limitation objection required a specific response in the circumstances; silence could not safely be interpreted as its rejection. In Hiro Balani v Spain, decided the same day, the court likewise examined a failure to address an argument of potential significance. The Court’s Guide on Article 6, civil limb, explains that reasoned decisions enable parties to understand that they have been heard and to exercise an available right of appeal.

This Convention doctrine governs its own legal setting; it is not directly applicable to Iranian courts or a universal mandate to write answers to every pleading point. Its transferable insight is narrower and more defensible: a court must make intelligible the path from the legally material disputes to its decision. Repetitive submissions need not be restated. Some objections become immaterial once a case is resolved on an independent ground. Other points can be answered by the judgment’s reasoning as a whole. The safeguard must respect these distinctions rather than turn completeness into a count of paragraphs.

A reverse-traceability test

I propose a short, judge-led reverse-traceability test. It begins not with the AI draft, but with the authorised record: the pleaded issues, material written submissions, relevant hearing record and evidence properly before the court. The judge identifies objections that, if well founded, would defeat an indispensable factual or legal step in the proposed disposition. The test then asks where the final reasons answer each such objection.

This is a counterfactual inquiry rather than a mechanical checklist. Suppose the party’s objection were correct. Could the order still stand on the grounds given? If yes, the reasons should make that independent ground discernible. If no, the judgment must address the disputed premise, apply the relevant rule, or explain why the objection fails. A reviewer must check whether the draft describes the submission fairly, not merely whether a keyword appears somewhere in the text.

The method can remain proportionate. A short civil dispute might contain one decisive objection; a complex appeal could contain several. No fixed number should be imposed. Nor should the judge’s initial list harden into a premature view of the case. If the preferred legal route changes, what was once immaterial may become crucial. The judge, not a model-generated ranking of “important issues”, must make that assessment and be able to revise it.

The proposed safeguard differs from routine citation verification in two ways. First, it tests completeness against a source independent of the draft. Secondly, it links completeness to legal consequence: what would change if the objection succeeded? It is not a proposal for an algorithm to determine outcomes. Nor does it assume that any existing AI product can reliably perform the task. The point is to specify what human review must accomplish even when the tool is fast and its prose plausible.

Oversight must be possible in practice

The judiciary’s October 2025 AI guidance warns against hallucinations and disclosure of confidential material, and emphasises the personal responsibility of judicial office holders for work produced in their names. These are indispensable cautions. But responsibility is meaningful only if the judge can inspect the underlying documents and challenge an automated account of the dispute. A procurement or deployment policy should preserve traceable access to sources; a polished interface that conceals the record can make meaningful verification more difficult.

That requirement has an institutional dimension. In controlled evaluations of authorised tools, courts could assess whether reviewers identify omitted outcome-sensitive objections, whether the final reasons treat them accurately, and how much time the procedure adds. Such testing should distinguish errors caused by document extraction, summarisation and drafting. It should also examine false alarms: a system that flags every minor disagreement as decisive could burden rather than protect the court. These are proposals for evaluation, not claims about measured performance.

Confidentiality supplies a hard boundary. Nothing here justifies feeding parties’ pleadings or non-public evidence into public AI services. The judiciary of England and Wales expressly confines judicial use to approved systems under its own arrangements. Other jurisdictions must apply their own rules on security, privilege, deliberative material and records retention. Where no approved tool exists, the same reverse-traceability check can be performed manually. The legal safeguard should not depend on purchasing technology.

Answering the strongest objections to the proposal

One concern is duplication: judges already read files and check drafts. The reply is that review needs a targeted object. An additional document or ritual is unjustified if the judge can perform the test in the ordinary course of settling reasons. A second concern is that the test might turn a flexible duty to give reasons into a demand for exhaustive judgments. That would be a mistake. The standard remains materiality, assessed in the light of the actual disposition. A precise sentence disposing of a pivotal point can be more valuable than pages of unnecessary authorities.

There is a third risk: a supposedly neutral list of decisive objections may itself reflect bias. The judge should therefore check the losing party’s strongest legally relevant case, not merely the arguments that fit an early draft. Institutional audits should ask whether the procedure consistently represents both sides. The proposed working aid also creates no new general right of public access to judicial notes or draft deliberations; those questions remain governed by each jurisdiction’s own law.

The measure of accountable judging

The strongest reason for the test is not administrative efficiency. It is the litigant’s ability to understand and challenge the court’s reasoning. A person who loses a case should be able to see why an outcome-sensitive objection failed, whether the reasons were typed by a judge or refined with an approved tool. A human signature beneath fluent text is not, by itself, evidence that the decisive point was considered.

The October 2026 position statement rightly insists on human oversight. The next task is to make that commitment operational. Before approving AI-assisted material, the judge should be able to move from the original objection, through the governing law and relevant evidence, to the reason actually adopted. If that path is broken, the drafting process is not finished. Judicial accountability begins where a confident summary stops.

Dr Ehsan Amirkhani is the presiding judge of Branch 2 of the General Civil Court of Kermanshah, Iran, and an independent legal researcher. He writes here in a personal academic capacity. No confidential case materials are used; all case examples are hypothetical.

Opinions expressed in JURIST Commentary are the sole responsibility of the author and do not necessarily reflect the views of JURIST's editors, staff, donors or the University of Pittsburgh.